AF: Small: Advances in Private Optimization

Modern Artificial intelligence (AI) systems are typically built with the use of large datasets that may be generated by individuals contributing data over the internet, such as through product ratings, comments, or other online interactions. As a result, it is critical that such AI systems are able to preserve the privacy of the individuals whose data is used: it should not be possible for an outside observer to learn anything about any individual person through the use of the AI system. This project will investigate the fundamental limits of the trade-off s between privacy and performance to build AI systems that are as performant as possible without compromising privacy. The project's results will not only improve the security of people who already benefit from products that learn from their data but will also reduce bias in AI by enabling more sensitive or vulnerable individuals to participate safely. On a technical level, this project will develop new differentially private stochastic optimization algorithms. In recent years, there has been a surge of interest in private optimization, but the theory for private non-convex optimization (which is required for training neural networks) is surprisingly underdeveloped. For this setting, the typical approach is to treat a standard non-private algorithm as a black box, providing it with inputs that have already been processed by adding noise to obscure individual contributions so that the output must also necessarily preserve privacy. This project will open up this black box to produce new algorithms that improve privacy guarantees while maintaining good convergence results both in theory and in practice. One of the initial technical approaches will be to develop a connection between momentum, a ubiquitous technique in modern optimization algorithms, with privacy. This approach does not require structural assumptions on the loss surface (e.g., convexity, smoothness) to ensure privacy but has the potential to significantly decrease the amount of noise injected into the algorithm, resulting in improved performance with the same level of privacy. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.